Small difference in predicted score when loading a trained Booster object into XGBClassifier

[protco]

Hi team,
We have a Booster trained using xgboost.train() and need to use it inside XGBClassifer() when doing score calibration. Here is the code, without explicitly setting n_classes_ and classes_, CalibratedClassifierCV complains that the model is not pre-fitted. With these settings, the predicted score using XGBClassifier is slightly different from the booster itself. Do you know any possible reasons and how to make the score 100% match?
xgb_classifier = XGBClassifier()
xgb_classifier.Booster = xgb_booster
xgb_classifier.n_classes = 2
xgb_classifier.classes_ = np.array([0, 1])
XGBoost version: 1.5.0
thank you,

[trivialfis]

Could you please:

• Update to the latest XGBoost.
• Provide a reproducible example if the issue persists with the latest version.

[protco]

Update: it turns out with our pre-trained booster loaded into XGBClassifier, XGBClassifier._Booster.predict() returns identical score as booster.predict(). So the difference we saw is actually between XGBClassifier._Booster.predict() and XGBClassifier.predict_proba(). The question is why do the two produce different result.

Thank you, Jiaming!
I tried xgboost 2.1.4 to load our pre-trained boosters, the prediction difference between the booster and XGBClassifier still remain the same (on average XBGClassifier prediction is ~1-5% lower for 3 different models). Unfortunately I can't share the booster files. I have tried both 1.5.0 and 2.1.4 on a toy model using the breast cancer dataset, and the predictions match perfectly for both versions. So not sure why this difference only happens with our pre-trained models. Are there any possible reasons I can test?

[trivialfis]

I missed the thread .. apologies. But then, I wouldn't be able to make any guess without seeing it myself.

[trivialfis]

Closing for now. Feel free to re-open if there's a reproducible example.